American Institute of Aeronautics and Astronautics

1

CFD Solver Comparison of Low Mach Flow over the

ROBIN Fuselage

Jennifer N. Abras1

NAVAIR Applied Aerodynamics and Store Separation Branch, Patuxent River, MD, 20670

and

Nathan Hariharan2

CREATE-AV, Lorton, VA, 22079

The ROtor Body INteraction (ROBIN) fuselage is used as a baseline test case for a

computational fluid dynamic (CFD) study of solver drag prediction results for low Mach

flow. All of the predictions are compared not only to one another, but also to available wind

tunnel data. Comparisons of integrated viscous and pressure drag, flow separation point,

and centerline pressure distributions are analyzed. Parametric studies of the independent

options available within each code for low Mach flow conditions are investigated. Grid

studies are also presented. The final comparisons reveal that for the attached flow regions

all of the CFD codes predict approximately the same result. The differences occur when the

flow begins to separate aft of the fuselage. Benefits are gained when the viscous grid layers

are merged from tetrahedrons into prisms, and when the incompressible option is employed.

Higher spatial order of accuracy in the separated region is found to slightly improve the

results.

I. Introduction

HE CREATE-AV program is focused on developing state-of-the-art software for use with both fixed wing and

rotary wing aircraft analysis. These codes are HPCMP CREATETM

-AV HELIOS for rotary wing analysis and

HPCMP CREATETM

-AV KESTREL for fixed wing analysis. Release versions of both of these codes are currently

in use in government, industry, and academia. However, the development process to expand the capabilities of these

codes and to improve on the present capabilities is ongoing. While these changes and improvements encompass a

wide range of topics, one specific area of improvement is the addition of a new near-body CFD solver in CREATE-

AV HELIOS.

Evaluation of major changes in the configuration of any CFD code are pursued through comparisons to both

available measured data as well as to the predictions produced by established CFD solvers. Planned changes to the

CREATE-AV HELIOS solver will be evaluated using these types of comparisons. The near-body CFD solver in the

current release version of CREATE-AV HELIOS, is NSU3D. Future versions of CREATE-AV HELIOS are

planned to transition to the kCFD solver which currently resides in CREATE-AV KESTREL. It is of interest to

compare both kCFD and NSU3D to better understand the possible impact that the transition will have on the

CREATE-AV HELIOS solver. Predictions from both of these codes will be compared not only to one another, but

also to available wind tunnel data and predictions from both the OVERFLOW and FUN3D solvers.

The wind tunnel test case selected for this comparison is centered around the ROBIN-mod7 isolated fuselage.

This test was run in the NASA Langley 2 ft x 3 ft wind tunnel. The purpose of this test is to investigate fuselage

drag reduction using aft blowing. However, the baseline cases run during this test are suitable for comparison

during the present study. Reference 1 contains a subset of the wind tunnel data from this test as well as the

OVERFLOW predictions. The outer mold line of this fuselage is defined by a series of equations that are provided

in appendix A of reference 1. All test data and OVERFLOW predictions presented herein have been digitized or

copied from this paper. Recent CREATE-AV HELIOS evaluations have also used this test data, comparing

CREATE-AV HELIOS to a commercial off the shelf (COTS) solver, CFD++2.

1 Aerospace Engineer, jennifer.abras@navy.mil, AIAA Professional Member.

2 Deputy Project Manager, nathan.hariharan.ctr@hpc.mil, AIAA Associate Fellow.

T

Dow

nloa

ded

by N

ASA

LA

NG

LE

Y R

ESE

AR

CH

CE

NT

RE

on

July

7, 2

015

| http

://ar

c.ai

aa.o

rg |

DO

I: 1

0.25

14/6

.201

4-07

52

52nd Aerospace Sciences Meeting

13-17 January 2014, National Harbor, Maryland

AIAA 2014-0752

This material is declared a work of the U.S. Government and is not subject to copyright protection in the United States.

AIAA SciTech

American Institute of Aeronautics and Astronautics

2

This paper compares the drag predictions of the CREATE-AV HELIOS solver and the CREATE-AV KESTREL

solver against available wind tunnel data and predictions made by OVERFLOW and FUN3D. These comparisons

include grid density studies, grid boundary layer analysis, incompressible solver evaluation, near-body trim distance

analysis, overall solver comparisons, and a timing and convergence comparison. The final comparisons utilize the

centerline pressure distributions, center slice velocity contours, flow separation location, and integrated pressure and

viscous drag coefficients in order to assess the performance of each code.

II. Methodology

A. HPCMP CREATETM

-AV HELIOS v4.0

CREATE-AV HELIOS3 is a multi-function code designed specifically for rotorcraft analysis. It contains

unstructured near-body and Cartesian off-body CFD solvers, overset methodology using PUNDIT, rigid grid

motion, and rotor blade structural deformations through RCAS or through prescribed deformations. The near-body

solver is NSU3D, a node-centered unstructured CFD code that solves the Reynolds Averaged Navier-Stokes

(RANS) equations applying 2nd

-order spatial accuracy. The Spalart-Allmaras (SA) and SA-DES turbulence models

are available. The unstructured grid can contain a mixture of tetrahedral and prismatic cells. ARC3D is the off-

body Cartesian solver which applies up to 5th

-order spatial accuracy. This solver may be run inviscid or with the

SA-DES turbulence model. This solver also has the option of applying automatic mesh refinement (AMR) to the

solution. Cases using NSU3D alone and dual NSU3D and ARC3D are presented. All cases apply the SA

turbulence model to the near-body, and the Euler solution in the off-body where applicable. All solutions are run

steady-state.

B. HPCMP CREATETM

-AV KESTREL v4.0.5

CREATE-AV KESTREL3 is a multi-function code designed specifically for fixed wing aircraft analysis. It

contains a near-body CFD solver (kCFD), control surface motion capability, engine analysis, and 6DoF analysis.

kCFD is a cell-centered unstructured CFD code which solves the RANS equation using 2nd

-order spatial accuracy.

A variety of turbulence models are available including the SA and SA-DDES models. The unstructured grid can

contain a mixture of tetrahedral and prismatic cells. This solver also has the option of applying automatic mesh

refinement (AMR) to the solution. All cases apply the SA turbulence model and are run steady-state.

C. FUN3D v12.2

The NASA Langley code FUN3D4

is an unstructured CFD code that solves the RANS equations using a node-

centered 2nd

-order upwind implicit scheme. The unstructured grid can contain a mixture of tetrahedral and prismatic

cells. This code is able to perform a variety of analyses including rigid and elastic grid motion, 6DoF analysis,

design optimization, and grid adaptation. This code also contains a variety of rotorcraft specific options including

coupling with external comprehensive rotorcraft codes. There are a wide variety of flux schemes and turbulence

models available, but for this study the Roe flux scheme and the SA turbulence model have been selected. All cases

are run steady-state.

III. Conditions

The wind tunnel model is shown in Figure 1. The

model is inverted in order to focus on the flow

separation on the lower aft portion of the fuselage. A

series of 41 pressure taps are located along the

centerline of the fuselage. These pressure taps run

along the length of the bottom of the fuselage, and

along the top of the fuselage between the nose and the

post. The flow conditions utilized include a Mach

number of 0.1 (about 34 m/s), a Reynolds number,

based on the fuselage length, of 1.6M and an angle of

attack of 0. Standard sea level conditions are

assumed in the CFD solutions. The mount shroud is

constructed using an extruded NACA0018 airfoil

section. Figure 1. Image of the wind tunnel model from Ref.1.

Dow

nloa

ded

by N

ASA

LA

NG

LE

Y R

ESE

AR

CH

CE

NT

RE

on

July

7, 2

015

| http

://ar

c.ai

aa.o

rg |

DO

I: 1

0.25

14/6

.201

4-07

52

American Institute of Aeronautics and Astronautics

3

IV. Grids

The underlying CAD used for this model is in the form of a NASTRAN file. This NASTRAN mesh data was

loaded into the Rhinoceros v4.0 CAD program where NURBS surfaces were generated. These NURBS surfaces

were used within the NASA Langley TetrUSS tools, GridTool and VGrid, to generate the tetrahedral unstructured

grids. All the solvers employed here use unstructured grids, the details for these grids are provided in Table 1. The

boundary layers are merged for a majority of the cases. This boundary layer merging is always performed using the

CREATE-AV HELIOS preprocessor. The resulting processed NSU3D grid is then used within CREATE-AV

KESTREL and FUN3D for consistency. For CREATE-AV KESTREL this grid is converted to AVMesh format

using the CREATE-AV KESTREL preprocessor, FUN3D is able to read in the NSU3D file format with no

modification. Some CREATE-AV

HELIOS cases were run with an

unstructured near-body and a Cartesian

off-body. This grid (Grid-5) contains a

constant cell size with trim distances

ranging from 0.25 inches to 0.5 inches.

The first set of cases run use the ROBIN-

mod7 fuselage with a pylon on the top as

shown in Figure 2. The wind tunnel model

has this removed in order to add the

mount. The corresponding “free-air”

OVERFLOW predictions were run

without this pylon. Thus, a second series

of cases were run for each solver using the

model with no pylon. The boundary layer

spacing in all cases is selected to ensure

that the y+ value is less than one over the

entire fuselage.

Table 1. Grids used during study.

Boundary Layer Cells # of Cells # of Nodes Description

Grid 1 Prisms 8.7M 2.3M Coarse grid case

Grid 2 Tetrahedra 24.0M 4.1M Refined grid case

Grid 2 Prisms 14.0M 4.1M Refined grid case

Grid 5 Prisms 15.5M 5.5M 0.05” cell size, HELIOS dual mesh only

Grid 6 Prisms 22.3M 5.3M No pylon, additional aft refinement

V. Results

All solutions were run using the Spalart-Allmaras turbulence model in order to compare more directly with the

available OVERFLOW data. This section looks at grid resolution and the impact of equation modifications before

performing any direct solver comparisons. It is important to note that the presence of the wind tunnel walls and

model mount do have an impact on the measured data. The CFD solutions presented here do not model the wind

tunnel walls or the mount. It was demonstrated that modeling the wind tunnel walls and mount is necessary in order

to match the test data in Reference 1. This comparison ran OVERFLOW with and without the tunnel modeled and

compared the solutions. The results with the tunnel are much closer to the experimental data than the results without

these features. Thus, a better comparison for the present study is achieved by using the OVERFLOW data

computed in “free-air” as the baseline.

A. Grid Independence Study

Grid-1, Grid-2, and Grid-5 were used to assess grid independence. Grid-5 does not have sufficiently far outer

boundaries and was designed specifically for the CREATE-AV HELIOS dual mesh capability. Figure 3 compares

CREATE-AV KESTREL solutions using Grid-1 and Grid-2. The data are nearly identical confirming that

approximate grid independence has been achieved for this case.

Figure 2. Surfaces used in the CFD grids with and without the

pylon.

Dow

nloa

ded

by N

ASA

LA

NG

LE

Y R

ESE

AR

CH

CE

NT

RE

on

July

7, 2

015

| http

://ar

c.ai

aa.o

rg |

DO

I: 1

0.25

14/6

.201

4-07

52

American Institute of Aeronautics and Astronautics

4

The CREATE-AV HELIOS grid comparisons are shown in Figure 4. The solutions reveal that for Grid-1 and

Grid-2 the solutions are nearly identical. This is also true of Grid-5 in the attached flow regions. However, there is

a large difference in the separated region when employing the dual mesh methodology with Grid-5. This is

hypothesized to be the result of having a finer off-body mesh in the region aft of the fuselage in the separated flow

region as well as having higher order spatial accuracy in the off-body solver. This hypothesis was partially tested by

creating an unstructured mesh (Grid-6) with additional refinement in the same aft region and running an isolated

NSU3D case on this mesh. The results do show some improvement in the separated region with the additional

refinement. It should be noted that the dual mesh case is not using the isolated NSU3D solver as is used for Grid-1

and Grid-2. The Grid-5 case uses a combined NSU3D/ARC3D solver where the ARC3D solver is applying the

Euler equations in the off-body.

-1

-0.5

0

0.5

1

1.5

2

0 0.2 0.4 0.6 0.8 1 1.2 1.4

Cp

x/R

Experiment TopHelios Grid 5Helios Grid 2Helios Grid 1

-1

-0.8

-0.6

-0.4

-0.2

0

0.2

0.4

0.6

0.8

1

0 0.2 0.4 0.6 0.8 1 1.2 1.4

Cp

x/R

Experiment BottomHelios Grid 5Helios Grid 2Helios Grid 1

Figure 4. Grid independence for CREATE-AV HELIOS solutions.

-1

-0.5

0

0.5

1

1.5

2

0 0.2 0.4 0.6 0.8 1 1.2 1.4

Cp

x/R

Experiment Top

Kestrel Grid 2

Kestrel Grid 1

-1

-0.8

-0.6

-0.4

-0.2

0

0.2

0.4

0.6

0.8

1

0 0.2 0.4 0.6 0.8 1 1.2 1.4

Cp

x/R

Experiment Bottom

Kestrel Grid 2

Kestrel Grid 1

Figure 3. Grid independence for CREATE-AV KESTREL solutions.

Dow

nloa

ded

by N

ASA

LA

NG

LE

Y R

ESE

AR

CH

CE

NT

RE

on

July

7, 2

015

| http

://ar

c.ai

aa.o

rg |

DO

I: 1

0.25

14/6

.201

4-07

52

American Institute of Aeronautics and Astronautics

5

B. Tetrahedral vs. Prismatic Boundary Layers

All the grids developed for this study use no-slip boundary conditions for the fuselage surfaces. This boundary

condition requires the presence of sufficient grid cells close to the surface in order to resolve the boundary layer

flow gradients. This is accomplished by layering high aspect ratio cells near the fuselage surfaces. These are often

called viscous layers. The grid generation software used for this study only generates tetrahedral cells. In these

cases, it is common practice to employ preprocessing software to convert the tetrahedral cell structure within the

viscous layers into a prismatic cell structure by combining, or merging, the tetrahedral cells. Solutions using the

unmerged viscous layer grids were computed using both CREATE-AV KESTREL and FUN3D. These solutions are

compared to the corresponding merged viscous layer grid solutions in Figure 5. Over the top of the fuselage there is

little difference between the four solutions. The structure of the boundary layer mesh starts to have an impact aft of

the pylon where there is a small region of flow separation. Over the bottom of the fuselage the boundary layer mesh

structure again has little impact over the front portion of the fuselage. However, in the region of separation there is a

noticeable difference between the solutions. This impact appears to be more significant for CREATE-AV

KESTREL than for FUN3D and may have to do with differences between the node-centered and cell-centered

methodologies. In general, the difference between the merged and the tetrahedral boundary layers is that the

prismatic cell faces are aligned perpendicular or normal to the surface whereas the tetrahedrons have faces that are

not all aligned with the surface. This impacts the computations in the boundary layer.

C. Compressible vs. Incompressible Solvers

FUN3D has the option to run either the compressible or the incompressible Navier-Stokes equations. These two

methods are compared in Figure 6. For the incompressible solution an artificial compressibility factor of 10.0 is

used. In the attached flow regions the solutions are identical. However, in the separated flow region the

compressible solution displays higher frequency fluctuations near the separation point while the incompressible

solution is much smoother and more closely follows the test data. This is not surprising since the compressible

equations can suffer from high equation stiffness in low Mach flows and as a result predict non-physical numerical

fluctuations in the solution. Typically, either a low Mach preconditioning method or switching to the

incompressible Navier-Stokes equations is employed to alleviate the stiffness problem. In this case, the numerical

fluctuations are directly impacting the separation region and thus the downstream predictions from this point.

-1

-0.5

0

0.5

1

1.5

2

0 0.2 0.4 0.6 0.8 1 1.2 1.4

Cp

x/R

Experiment TopFUN3D Grid 2FUN3D Grid 2 TetsKestrel Grid 2Kestrel Grid 2 tets

-1

-0.8

-0.6

-0.4

-0.2

0

0.2

0.4

0.6

0.8

1

0 0.2 0.4 0.6 0.8 1 1.2 1.4

Cp

x/R

Experiment BottomFUN3D Grid 2FUN3D Grid 2 TetsKestrel Grid 2Kestrel Grid 2 tets

Figure 5. Centerline pressures comparing tetrahedral vs. prismatic boundary layers.

Dow

nloa

ded

by N

ASA

LA

NG

LE

Y R

ESE

AR

CH

CE

NT

RE

on

July

7, 2

015

| http

://ar

c.ai

aa.o

rg |

DO

I: 1

0.25

14/6

.201

4-07

52

American Institute of Aeronautics and Astronautics

6

A low Mach number preconditioning option is available in CREATE-AV HELIOS and was pursued as part of

this study. These cases investigated changes to the βmin parameter. However, converged cases were not obtainable

and thus will not be presented here.

D. Comparison of Near-Body Trim Distance in HELIOS Dual Mesh Cases

The dual mesh feature in CREATE-AV HELIOS employs a near-body unstructured mesh surrounded by a

Cartesian off-body mesh. The position of the overlapping region between these meshes is adjusted by trimming the

near-body mesh to a fixed distance from the solid surfaces. This method cuts away near-body cells that lie beyond

the user defined trim distance before the off-body mesh is generated. A set of cases was run to evaluate the impact

of the near-body trim distance on the solution. This case is set-up specifically so that the only variable that changes

is the position of the near-body and off-body mesh fringes. This was accomplished by generating a near-body grid

with a constant cell size throughout the mesh (Grid-5). The exception to this constant cell size is the viscous grid

layering which use the same

surface normal size distribution

as the single mesh cases.

Around this near-body mesh is a

fixed off-body Cartesian mesh

with the same constant cell size

that is set in the near-body. The

near-body mesh trim distances

applied are 0.25 in, 0.3 in, 0.35

in, and 0.5 in. A sample slice

through the 0.3 in mesh is

provided in Figure 7.

There are two considerations

when selecting a near-body trim

distance; geometric and solution

based. The geometric issues

involve the quality of the fringes

supplied to PUNDIT to overset

the near-body and off-body

meshes. The solution based

issues involve making sure that

-1

-0.5

0

0.5

1

1.5

2

0 0.5 1

Cp

x/R

Experiment TopFUN3D Grid 2 IncompFUN3D Grid 2

-1

-0.8

-0.6

-0.4

-0.2

0

0.2

0.4

0.6

0.8

1

0 0.5 1

Cp

x/R

Experiment Bottom

FUN3D Grid 2 Incomp

FUN3D Grid 2

Figure 6. Centerline pressures comparing compressible and incompressible solutions in FUN3D.

Figure 7. Grid slice through the dual mesh for the 0.3 in trim case (red =

near-body, blue = off-body).

Dow

nloa

ded

by N

ASA

LA

NG

LE

Y R

ESE

AR

CH

CE

NT

RE

on

July

7, 2

015

| http

://ar

c.ai

aa.o

rg |

DO

I: 1

0.25

14/6

.201

4-07

52

American Institute of Aeronautics and Astronautics

7

-1

-0.5

0

0.5

1

1.5

2

0 0.2 0.4 0.6 0.8 1 1.2 1.4

Cp

x/R

Experiment TopHelios Grid 2 N/AHelios Grid 5 0.5Helios Grid 5 0.35Helios Grid 5 0.3

-1

-0.8

-0.6

-0.4

-0.2

0

0.2

0.4

0.6

0.8

1

0 0.2 0.4 0.6 0.8 1 1.2 1.4

Cp

x/R

Experiment BottomHelios Grid 2 N/AHelios Grid 5 0.5Helios Grid 5 0.35Helios Grid 5 0.3

Figure 9. Centerline pressures comparing near-body trim distance in Helios dual mesh cases.

the Euler off-body solution does not cover any regions where viscous effects are important. The consequences of

exceeding the geometric limit are best shown in the 0.25 in trim case. This case computed a total of 4515 orphans.

These orphans are illustrated in Figure 8 where the orphan locations are represented by the green spots. All of these

orphans are located in places where the off-body Cartesian mesh intersects the second viscous grid layer. In this

plot the black represents the off-body mesh, the red and blue represent the near-body mesh where the blue is the

near-body fringe. Therefore, the trim distance must be increased in order to move the off-body fringe away from the

viscous grid layers and thus reduce the number of orphans. The 0.3 in case has a total of 241 orphans. This is below

the default reject threshold of 1000 and was allowed to complete. The 0.35 in and the 0.5 in cases both compute

zero orphans.

The next step is to make sure that the Euler solution does not overlap the viscous regions. This is more difficult

to identify, but generally requires that the off-body not cut into the boundary layer. The centerline pressure

predictions for the 0.3 in, 0.35 in, and the 0.5 in trim cases are compared in Figure 9. It is apparent from this

comparison that the Euler assumptions in the off-body are interfering with the near-body predictions in the 0.3 in

case. This is further illustrated in Figure 10 where there is additional unsteadiness predicted around the attached

flow regions of the fuselage. The solutions for the 0.35 in and the 0.5 in cases are much smoother and predict the

same solution for the attached flow regions. Thus, in these regions the Euler off-body is not overlapping the

Figure 8. Orphan locations for the 0.25 in trim distance case.

Dow

nloa

ded

by N

ASA

LA

NG

LE

Y R

ESE

AR

CH

CE

NT

RE

on

July

7, 2

015

| http

://ar

c.ai

aa.o

rg |

DO

I: 1

0.25

14/6

.201

4-07

52

American Institute of Aeronautics and Astronautics

8

boundary layer for trim distances greater than 0.35 in.

In the separated flow region, the solutions begin to diverge. The single mesh case is included to show that as the

trim distance is increased the solution is trending towards the single mesh case. However, the equivalent single

mesh case would actually use a much finer grid than the Grid-2 case plotted here. There seems to be an optimum

point between 0.3 in and 0.35 in that would produce the closest solution. However, a solution dependent on trim

distance is not the ideal situation. The off-body mesh is intended to provide both efficiency and accuracy to the

solution through quicker per node computation time, increased spatial order of accuracy, and adaptive mesh

refinement. In this case, for which the mesh refinement is fixed, the variables that differ between the near-body and

the off-body are primarily the inviscid flow assumption and the increased spatial accuracy. The spatial accuracy

would certainly increase the resolution of the flow structures in the flow aft of the fuselage in the off-body region. It

may be that the increased spatial order of accuracy close to the surface is impacting the separation characteristics, or

it may be that the trim distance is not sufficiently large in the separated flow region so that the Euler flow

assumption is valid.

E. Initial Solver Comparison

This initial solver comparison looks specifically at comparing the best solutions obtained with each solver using

the geometry with the pylon included. For CREATE-AV KESTREL this is the Grid-2 merged boundary layer case,

for FUN3D this is the Grid-2 merged boundary layer incompressible case, and for CREATE-AV HELIOS this is the

Grid-5 combined NSU3D/ARC3D

solution with the additional off-

body refinement aft of the fuselage

and the 0.35 in trim distance. The

isolated NSU3D solution using

Grid-2 is included since it matches

the FUN3D and CREATE-AV

KESTREL conditions the closest.

These solutions are plotted in

Figure 11. For the attached flow

region the solutions are almost

identical, there is a small

divergence for the NSU3D solution

along the bottom of the fuselage.

In the separated flow regions each

code predicts a different solution.

The CREATE-AV HELIOS dual

mesh and CREATE-AV

KESTREL solutions seem to

predict the peak the best, while

FUN3D seems to predict the

separation point the best. The

isolated NSU3D solution has the

greatest error. It should be noted

that since the wind tunnel walls are

not modeled here that it is

unknown exactly how far off the

free air predictions should be from

the measured data. The next

section that compares the

OVERFLOW data with the current

predictions will approximate this

delta.

0.30”

0.35”

0.50”

Figure 10. Centerline slices comparing the velocity contours for

different near-body trim distances.

Dow

nloa

ded

by N

ASA

LA

NG

LE

Y R

ESE

AR

CH

CE

NT

RE

on

July

7, 2

015

| http

://ar

c.ai

aa.o

rg |

DO

I: 1

0.25

14/6

.201

4-07

52

American Institute of Aeronautics and Astronautics

9

The comparison of the oil line plots with the test data in Figure 12 reveal somewhat different conclusions. In

these plots, FUN3D and CREATE-AV HELIOS dual mesh predict the separation point the closest, while CREATE-

AV KESTREL and NSU3D are separating later. The red circles indicate the experimental separation point.

The integrated drag listed in Table 2 indicates that each code is predicting a somewhat different drag coefficient.

This is true of both the viscous and the pressure drag components. It should be noted that the experimental value is

expected to be higher than the computed values since the tunnel walls are not present in the CFD solutions.

Figure 12. Oil line plots comparing flow separation points against experiment.

-1

-0.5

0

0.5

1

1.5

2

0 0.2 0.4 0.6 0.8 1 1.2 1.4

Cp

x/R

Experiment TopFUN3D Grid 2 IncompKestrel Grid 2Helios Grid 2Helios Grid 5

-1

-0.8

-0.6

-0.4

-0.2

0

0.2

0.4

0.6

0.8

1

0 0.2 0.4 0.6 0.8 1 1.2 1.4

Cp

x/R

Experiment Bottom

FUN3D Grid 2 Incomp

Kestrel Grid 2Helios Grid 2

Helios Grid 5

Figure 11. Centerline pressures comparing the best solutions for the model with the pylon included.

Dow

nloa

ded

by N

ASA

LA

NG

LE

Y R

ESE

AR

CH

CE

NT

RE

on

July

7, 2

015

| http

://ar

c.ai

aa.o

rg |

DO

I: 1

0.25

14/6

.201

4-07

52

American Institute of Aeronautics and Astronautics

10

F. Final Solver Comparison

This final comparison provides the opportunity to compare the three current solvers against the OVERFLOW

data, since the same geometry with the pylon removed is used for all of these cases. The best solution for each code

is compared in this section. For CREATE-AV HELIOS and for CREATE-AV KESTREL these are the standard

compressible case, and for FUN3D this is the incompressible case.

Table 3 lists the integrated drag for each code.

In this table the OVERFLOW solutions include a

version with the tunnel walls modeled and a

version that does not include the tunnel walls in

the model. The tunnel solution is expected to

match the experimental data the closest, and in

fact the drag coefficients are the same, while the

free-air solution is used to represent the expected

delta between the experimental data and the

remaining CFD solutions. Thus, the percent

errors listed in the last column are computed

against the OVERFLOW free-air data. Each of

the three solvers predicts a different integrate

drag. Both CREATE-AV HELIOS and

CREATE-AV KESTREL are underpredicting the

drag coefficient, but to different degrees. Both

codes seem to be underpredicting the pressure

drag when compared with the OVERFLOW data,

while CREATE-AV KESTREL is the closest to

the viscous drag. FUN3D is overpredicting the

drag, where the viscous drag is lower than the

OVERFLOW free-air data while the pressure drag

is much higher. The viscous drag is dependent on

the gradients close to the surface, while the

pressure drag is going to vary depending on the

separation point aft of the fuselage. Looking at

Table 3. Integrated drag separated into pressure and viscous components (%error compared to

OVERFLOW free-air results).

Drag

Coefficient

Viscous Drag

Coefficient

Pressure Drag

Coefficient Percent Error

Experiment 0.145

OVERFLOW (Tunnel) 0.145 0.055 0.090

OVERFLOW (Free-air) 0.114 0.056 0.058

HELIOS Grid-6 0.082 0.049 0.034 -27.7%

KESTREL Grid-6 0.108 0.057 0.051 -5.49%

FUN3D Grid-6 (Incomp) 0.120 0.053 0.067 5.56%

Table 2. Integrated drag subdivided into viscous and pressure components.

Drag

Coefficient

Viscous Drag

Coefficient Pressure Drag Coefficient

Experiment 0.145

KESTREL Grid-2 0.114 0.059 0.055

HELIOS Grid-5 (Dual Mesh) 0.088 0.044 0.044

HELIOS Grid-2 0.086 0.050 0.036

FUN3D Grid-2 (Incomp) 0.129 0.056 0.073

FUN3D Grid 6Incompressible

Kestrel Grid 6

Helios Grid 6

Figure 13. Oil line plots comparing flow separation

points against experiment for Grid-6.

Dow

nloa

ded

by N

ASA

LA

NG

LE

Y R

ESE

AR

CH

CE

NT

RE

on

July

7, 2

015

| http

://ar

c.ai

aa.o

rg |

DO

I: 1

0.25

14/6

.201

4-07

52

American Institute of Aeronautics and Astronautics

11

FUN3D Grid 6 Incompressible

Helios Grid 6

Kestrel Grid 6

Figure 14. Centerline slices comparing the velocity

contours.

the distributed predictions will clarify the

differences seen in the table.

Figure 13 plots the oil line plots in the

separation region aft of the fuselage for CREATE-

AV HELIOS, CREATE-AV KESTREL, and

FUN3D compared to the tunnel data. The

separation point in FUN3D is the earliest and

corresponds with the higher pressure drag seen in

the table. The CREATE-AV KESTREL separation

point is next followed by the CREATE-AV

HELIOS separation point. The corresponding

pressure drag components also follow this trend.

The increased grid density aft of the fuselage is

having a positive effect on the solution compared to

the Grid-2 cases plotted in the previous section.

The centerline slice contours are plotted in Figure

14 for comparison. These plots indicate that the

three codes have similar, but not identical separated

flow characteristics. The integrated pressure drag

indicates that the OVERFLOW solution separates

after FUN3D, but before CREATE-AV HELIOS

and CREATE-AV KESTREL. The separation data

are not available to confirm this, but the pressure

tap data are available for comparison.

The centerline pressures plotted in Figure 15 also include the OVERFLOW data with and without the tunnel

walls. The OVERFLOW solution with the tunnel walls matches the experimental data the closest. When these

walls are removed the magnitude of the pressure coefficient drops. All of the CFD solutions without the tunnel

walls match in the attached flow regions along the top of the fuselage and along the bottom for the front half of the

fuselage. In the separated region the solutions begin to diverge. The general trends are the same, but the

magnitudes are different. CREATE-AV KESTREL matches the OVERFLOW data the closest. FUN3D matches

the test data after the separation point the closest. The CREATE-AV HELIOS (NSU3D) data are the farthest from

the predictions. A matching dual mesh case was not run without the pylon, but based on earlier results it is expected

that this would improve the CREATE-AV HELIOS prediction.

-1

-0.5

0

0.5

1

1.5

2

0 0.2 0.4 0.6 0.8 1 1.2 1.4

Cp

x/R

Experiment TopFUN3D Grid 6 IncompKestrel Grid 6Helios Grid 6Overflow Free-AirOverflow Tunnel

-1

-0.8

-0.6

-0.4

-0.2

0

0.2

0.4

0.6

0.8

1

0 0.2 0.4 0.6 0.8 1 1.2 1.4

Cp

x/R

Experiment Bottom

FUN3D Grid 6 Incomp

Kestrel Grid 6

Helios Grid 6

Overflow Free-Air

Overflow Tunnel

Figure 15. Centerline pressures comparing the best solutions for the model with the pylon removed.

Dow

nloa

ded

by N

ASA

LA

NG

LE

Y R

ESE

AR

CH

CE

NT

RE

on

July

7, 2

015

| http

://ar

c.ai

aa.o

rg |

DO

I: 1

0.25

14/6

.201

4-07

52

American Institute of Aeronautics and Astronautics

12

Overall, the best solutions from all four CFD codes are consistent with one another except for the separated flow

region where the character of the solutions are the same, but the magnitudes differ. These differing magnitudes in

the predicted results are impacting the integrated drag. Overall, the isolated NSU3D results are not performing as

well as the isolated kCFD results.

G. Computational Metrics

The computational metrics give a good idea of how fast each code is able to run. However, the timing alone

does not always give the full picture. This section compares both the computational time used as well as the

solution convergence characteristics to provide a full picture of how fast a solution may be expected to complete.

1) Timing Comparison

All cases presented here were run on Spirit. Spirit is an SGI ICEX machine running on Intel E5 Sandy Bridge

2.6 GHz processors with 16 cores per node and 32Gb per node. Some variability in the run time is detected when

the timing of similar restart cases is compared. Thus, the timing comparisons provided in Table 4 cannot be

considered absolute, but do allow for general code to code comparisons. The times listed in the table also include all

of the pre- and post-processing steps included in the case. This was done for consistency since internal time per

iteration values are not available for all codes. Instead the total time for each case is used in the table. When

comparing the results, the CREATE-AV HELIOS dual mesh methodology runs the fastest in terms of the time per

processor per iteration per degree of freedom. This is followed by NSU3D then CREATE-AV KESTREL, with

FUN3D as the slowest of the codes.

2) Convergence Comparison

The convergence characteristics tell the opposite story. Figure 16 compares plots of the drag coefficient

convergence for each of the cases listed in Table 4. In these plots FUN3D converges the fastest followed by

CREATE-AV KESTREL then CREATE-AV HELIOS dual mesh with NSU3D as the slowest to converge. The x-

axis of each of these plots represents the total iterations that the code was run for each of these cases. The isolated

NSU3D case ran up to one million iterations. This case was run past this point, by renumbering the restart directory

to a smaller iteration number, but because of internal formatting limits the iterations beyond one million were not

saved. The CREATE-AV HELIOS dual mesh case did not reach a single steady-state value because the mesh

behind the fuselage is much finer than for the Grid-2 cases. Thus, the refinement of the unsteady flow in this region

prevents the solution from reaching a single value. The FUN3D and CREATE-AV KESTREL cases converged

fairly quickly and were run well past the minimum convergence point.

In order to get a better idea of how long a user would expect to wait for a solution to complete, the minimum

convergence points are identified and extracted from the convergence plots. Obtaining a solution overnight is used

as the ideal situation; this represents a wall time of less than 12 hours. The minimum convergence points are

identified using the drag convergence only. In reality, the moments take longer to converge, but the drag

convergence may be used for comparison purposes. These minimum values are listed in Table 5. By extracting the

time per iteration from the previous table and multiplying by the minimum required iterations, a minimum total wall

clock time may be estimated. This wall time represents how long a user would have to wait to obtain a completed

solution.

Comparing these values reveals an interesting reversal. Even though CREATE-AV HELIOS runs the fastest, it

converges the slowest, requiring many more iterations to complete. Thus, the wall clock time does not allow for

Table 4. Computational timing comparisons.

Code(s) Run Grid Degrees of Freedom Procs

Wall

Clock

[hrs]

Iterations Time/proc/it/DoF

[nanosec]

HELIOS

(NSU3D & ARC3D) Grid-5

55M (nodes near)

91M (nodes off) 512 2.77 10K 0.0203

HELIOS

(NSU3D) Grid-2 4.1M nodes 512 1.95 50K 0.0673

KESTREL Grid-2 14.0M cells 512 24.39 50K 0.2441

FUN3D

(Incompressible) Grid-2 4.1M nodes 256 5.19 30K 0.5971

Dow

nloa

ded

by N

ASA

LA

NG

LE

Y R

ESE

AR

CH

CE

NT

RE

on

July

7, 2

015

| http

://ar

c.ai

aa.o

rg |

DO

I: 1

0.25

14/6

.201

4-07

52

American Institute of Aeronautics and Astronautics

13

obtaining a solution overnight under the test conditions. It should be noted that NSU3D does have a multi-grid

option which should improve the convergence, but this option was not attempted during this study. CREATE-AV

KESTREL and FUN3D both converge much faster allowing for the possibility of obtaining a solution overnight

even though the codes do not run as fast as CREATE-AV HELIOS.

VI. Conclusion

A study of three locally available CFD solvers (HPCMP CREATETM

-AV HELIOS, HPCMP CREATETM

-AV

KESTREL, and FUN3D) are compared against both experimental and OVERFLOW data obtained from reference 1.

These three solvers are compared using variations in the grid and solver parameters. The conclusions from this

study include,

Grid independence studies comparing Grid-1 and Grid-2 in both CREATE-AV KESTREL and CREATE-AV

HELIOS indicate that grid independence has been reached in both solvers.

Table 5. Estimated computational timing comparisons.

Code(s) Run Grid Degrees of Freedom

Estimated

iterations to

converge

Procs

Estimated

Wall Time

[hrs]

HELIOS

(NSU3D & ARC3D) Grid-5

5.5M (nodes near)

91M (nodes off) 100K 512 27.70

HELIOS

(NSU3D) Grid-2 4.1M nodes 1M 512 39.00

KESTREL Grid-2 14.0M cells 9K 512 4.39

FUN3D

(Incompressible) Grid-2 4.1M nodes 2.5K 256 0.43

Figure 16. Drag coefficient convergence comparison for selected cases (x-axis maximum represents total

iterations run).

Dow

nloa

ded

by N

ASA

LA

NG

LE

Y R

ESE

AR

CH

CE

NT

RE

on

July

7, 2

015

| http

://ar

c.ai

aa.o

rg |

DO

I: 1

0.25

14/6

.201

4-07

52

American Institute of Aeronautics and Astronautics

14

The comparison of tetrahedral vs. prismatic boundary layers indicates that the surface aligned faces in the

prismatic boundary layer yield better results than the unaligned faces in the tetrahedral boundary layer. This

difference is stronger in the cell-centered CREATE-AV KESTREL code than in the node-centered FUN3D

code.

The use of the incompressible equations reduces the incidence of non-physical numerical fluctuation in the

FUN3D solution.

When setting up a dual mesh case in CREATE-AV HELIOS care must be taken when defining the fringe

distance to both create a proper fringe for the oversetting process, as well as to keep sufficient distance of the

off-body solver from the viscous flow regions.

The initial comparison with the pylon included indicate that in the attached flow regions all of the solutions

predict approximately the same results. The separated regions predict different magnitude solutions with the

isolated NSU3D solution being the furthest from the experimental data.

The final comparison with the pylon removed allows for the comparison with the OVERFLOW data. In this

comparison all of the local CFD solvers predicted the drag to be within 6% error with the exception of the

CREATE-AV HELIOS prediction. The distributed data showed that the differences are primarily in the

separated flow aft of the fuselage.

The computational metrics show that CREATE-AV HELIOS runs the fastest and CREATE-AV KESTREL and

FUN3D run the slowest. However, the increased rate of convergence in CREATE-AV KESTREL and FUN3D

allow for a solution to be obtained overnight, while more time is needed for a CREATE-AV HELIOS solution.

Finally, it should be pointed out that this study did not do a thorough investigation of optimal solver settings,

default settings in the solvers were used in most cases, only a few deviations from these settings were attempted.

Since HELIOS is applied primarily to rotor-based flows, as opposed to isolated fuselages like the fixed-wing kCFD

code, the default settings in HELIOS may not have been ideal for this fuselage case. Future work will investigate

alternative solver settings in HELIOS in order to see if the results improve.

Acknowledgments

The authors would like to acknowledge the support of the HPCMP CREATETM

-AV Program and the

supercomputing resources provided by the HPCMP, in particular the Air Force Research Lab (AFRL). The

contribution of Rajneesh Singh of providing the original ROBIN surface mesh is also gratefully acknowledged.

References

1Schaeffler, N. W., Allan, B. G., Lienard, C., and Le Pape, A., “Progress Towards Fuselage Drag Reduction via

Active Flow Control: A Combined CFD and Experimental Effort,” 36th

European Rotorcraft Forum, Paris, France,

September 7-9, 2010.

2Singh, R., and Dinavahi, S., “Aerodynamic Force Computations for Rotorcraft Fuselage,” 50

th AIAA Aerospace

Sciences Meeting, Nashville, TN, January 9-12, 2012.

3HPCMP CREATE

TM-AV On-line Resources, “https://portal.create.hpc.mil/authdocs/av/index.php,” Last Accessed

November 18th

, 2013.

4FUN3D On-line Resources, “http://fun3d.larc.nasa.gov,” Last Accessed November 18

th, 2013.

Dow

nloa

ded

by N

ASA

LA

NG

LE

Y R

ESE

AR

CH

CE

NT

RE

on

July

7, 2

015

| http

://ar

c.ai

aa.o

rg |

DO

I: 1

0.25

14/6

.201

4-07

52